Long short-term memory (LSTM) deep learning-based prediction of sensor performance time series for improved VIIRS reflective solar band calibration

Past experience from postlaunch calibrations of Suomi-NPP (SNPP), NOAA-20, and NOAA-21 Visible Infrared Imaging Radiometer Suite (VIIRS) indicates the critical need for quick and accurate predictions of sensor optical throughput response changes. For example, the solar diffuser is a critical component of the VIIRS instrument and serves as a reference standard for the on-orbit calibration of VIIRS reflective solar bands (RSBs). However, the solar diffuser can experience degradation over time resulting from various factors, including exposure to solar ultraviolet (UV) radiation, energetic particles, and contaminants in the space environment. The changes in the optical properties of the solar diffuser material can impact the accuracy and stability of VIIRS radiometric calibration. SNPP VIIRS RSB suffered rapid postlaunch optical throughput degradation in the near-infrared (NIR) band gains (inverse of solar-F factor calibration coefficient) due to mirror contamination along the optical path. Given the short-term and long-term VIIRS radiometric calibration update needs and the delays between planning and execution of post-launch calibration updates, it is critical to accurately predict VIIRS sensor response changes in days or weeks ahead. The advancement of recurrent neural network (RNN) machine learning algorithm for time series prediction provides potential means for fast and accurate calibration time series prediction. In this study, long short-term memory (LSTM) RNN model is used to train and predict VIIRS calibration time series such as VIIRS solar diffuser spectral reflectance change time series. LSTM neural networks are a type of RNN architecture designed to model sequential data and capture short- and long-term dependencies and are well-suited for calibration time series forecasting due to their ability to remember information over extended time periods and to handle sequences of varying lengths. The calibration time series of solar diffuser reflectance change from SNPP VIIRS are used as example input sequences and target values to train the LSTM model. The prediction performance is assessed in terms of prediction error as a function of prediction horizons from one to seven days for spectral bands of VIIRS. The relative Root Mean Square Error (RMSE) of seven-day LSTM predictions of spectral degradations of SNPP solar diffuser reflectance is within 0.2%. The suitability of applying LSTM machine learning model for VIIRS calibration time series predictions is discussed.